KRK.FLIGHTS / EPKK
Our fog model: how it works
We train a model to recognise when visibility at Kraków Airport will fall below 550 metres two hours after the latest observation. It learned from the airport’s history. We now check its forecasts against new observations.
Still being tested. The timeline uses observations, the airport forecast and the area forecast. Our fog model does not yet change assessments or notifications.
What does the model look for?
- Moisture and cooling. It checks the gap between temperature and dew point, the temperature at which moisture starts to condense. It also considers whether that gap is closing.
- Wind and visibility. It considers wind speed, the latest visibility and changes across recent readings. One observation cannot describe the whole weather trend.
- Time of day, season and recent weather. The hour, month and recent observations, including precipitation, help distinguish an autumn night from a summer afternoon.
What did it learn from?
On 105,029 examples from 2018–2023, prepared from EPKK METAR observations. Each example connects earlier weather with visibility measured later. A prediction uses only information already available at that time.
Separate 2024 data were used to adjust predicted probabilities to the observed event frequency. This is a different stage from training on earlier years. Testing against new observations checks whether performance holds today.
Why use Kraków Airport’s own data?
Airport weather can differ from central Kraków. Terrain, night-time cooling and warmer urban areas contribute to those differences. The airport’s history allows local patterns to be learned without assuming identical weather across the city.
We separately test extra information about the sun’s position, cooling rate and wind direction. More data do not always improve a model. We do not automatically increase the assessment because the airport lies in a valley or near a city.
The live test: what do we know?
We compare our model, the airport forecaster’s prediction and a simple forecast that keeps the current conditions unchanged. All three are checked against the same later observations. Good-visibility hours alone are not enough: we also need cases of worsening visibility.
The test is just starting. We are waiting for forecasts saved before their target time and later observations that allow all three sources to be compared.
The table appears after 30 paired comparisons. This is a display threshold, not evidence of accuracy. Evaluating the model requires a larger sample, foggy days and different seasons.
How are the results calculated?
Forecast error is the Brier score: it compares predicted probabilities with later observations. Zero means perfect forecasts; a score of 0.01 does not mean 99% accuracy. False alarms and misses use a 30% probability threshold for visibility below 550 m. We count weather observations, not cancelled flights or sent alerts.
The target is two hours after the latest observation. If that observation is already 30 minutes old, 90 minutes remain. Verification must be within 15 minutes of the target time. Successive observations from the same fog episode are not independent cases.
The airport forecast does not always assign a numeric probability to a change. When such temporary or gradual changes affect the event being evaluated, we exclude that case from the paired comparison. Missing data do not become a good-weather forecast.
When will it help on the timeline?
When it demonstrates an advantage over the airport forecast and unchanged-weather prediction. We check both missed poor conditions and unnecessary warnings, especially during fog and across seasons. A short run of good results is not enough.
The model predicts general visibility, not runway visibility. The 550 m threshold does not determine whether a particular aircraft can land or the chance of a cancellation. See how weather is assessed →
Separate evaluation of the airport forecast
We also evaluate our interpretation of the airport forecast for visibility below 1000 m. This is a separate evaluation with a different threshold and question from the 550 m fog-model test. It does not measure flight-status accuracy.
Matched forecasts: 0. With a numeric visibility forecast: 0.